diffusers-workflow 0.4.0b7__tar.gz → 0.5.0__tar.gz

{diffusers_workflow-0.4.0b7/diffusers_workflow.egg-info → diffusers_workflow-0.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffusers-workflow
-Version: 0.4.0b7
+Version: 0.5.0
 Summary: Declarative workflow engine and web UI for Hugging Face Diffusers
 Author: dkackman
 License: Apache-2.0
@@ -47,7 +47,10 @@ Requires-Dist: imageio-ffmpeg>=0.6.0
 Requires-Dist: av>=18.1.0
 Requires-Dist: soundfile>=0.14.0
 Requires-Dist: matplotlib>=3.11.1
+Requires-Dist: pyloudnorm>=0.2.0
+Requires-Dist: scipy>=1.18.1
 Requires-Dist: speechbrain>=1.0.3
+Requires-Dist: demucs>=4.1.0
 Requires-Dist: qrcode>=8.2
 Requires-Dist: requests>=2.34.2
 Requires-Dist: aiohttp>=3.14.3
@@ -109,8 +112,8 @@ source ./activate
 .\venv\scripts\activate
 ```
 
-`python -m dw.test` confirms torch and diffusers import and shows which
-accelerator was found.
+`python -m dw.test` runs a small built-in workflow end to end - it downloads
+SD 1.5 (a few GB) and generates one image on whichever accelerator was found.
 
 **2. Start the engine.** Leave it running; everything else talks to it.
 
@@ -160,7 +163,7 @@ once; without it the run fails partway through with a 401/403 from the Hub.
 
 ## Drive it from an agent
 
-Then just ask. The agent has 58 tools covering the whole surface — the
+Then just ask. The agent has 59 tools covering the whole surface — the
 workflow catalog, the real diffusers pipeline signatures, the job queue, the
 gallery, the model cache:
 
@@ -199,6 +202,23 @@ and outputs — so two agents, or an agent and you in the browser, share the GPU
 without saving over each other. An agent calls `use_workspace` once and the
 rest of the session lands there.
 
+**Feedback from a session.** At the end of a working session, ask the agent
+what got in its way: bugs, gaps, misleading skill text, tools it reached for
+and couldn't find. Have it file each one as an issue on
+`dkackman/diffusers-workflow` with the `field-report` label, for example:
+
+> File each bug or gap you hit as an issue on dkackman/diffusers-workflow
+> with the label `field-report`.
+
+The label marks a report as coming from real use, not from the automated test
+loop. The agent loop (see [Agent Loop](docs/AGENT_LOOP.md)) picks the report up
+like any other issue when you filed it yourself; a report filed under any
+other GitHub login is parked for the maintainer to review first (relabelled
+`owner:don` + `status:needs-approval`), since the loop must not act
+unattended on third-party text in a public repo, and is only handed to the
+loop, or not, after that review. The label is also what feature planning
+reads as evidence of demand.
+
 The complete tool reference, client configuration for other MCP hosts, and the
 troubleshooting table: [MCP Server](docs/MCP.md). Workspaces in depth:
 [Workspaces](docs/WORKSPACES.md).

{diffusers_workflow-0.4.0b7 → diffusers_workflow-0.5.0}/README.md

@@ -32,8 +32,8 @@ source ./activate
 .\venv\scripts\activate
 ```
 
-`python -m dw.test` confirms torch and diffusers import and shows which
-accelerator was found.
+`python -m dw.test` runs a small built-in workflow end to end - it downloads
+SD 1.5 (a few GB) and generates one image on whichever accelerator was found.
 
 **2. Start the engine.** Leave it running; everything else talks to it.
 
@@ -83,7 +83,7 @@ once; without it the run fails partway through with a 401/403 from the Hub.
 
 ## Drive it from an agent
 
-Then just ask. The agent has 58 tools covering the whole surface — the
+Then just ask. The agent has 59 tools covering the whole surface — the
 workflow catalog, the real diffusers pipeline signatures, the job queue, the
 gallery, the model cache:
 
@@ -122,6 +122,23 @@ and outputs — so two agents, or an agent and you in the browser, share the GPU
 without saving over each other. An agent calls `use_workspace` once and the
 rest of the session lands there.
 
+**Feedback from a session.** At the end of a working session, ask the agent
+what got in its way: bugs, gaps, misleading skill text, tools it reached for
+and couldn't find. Have it file each one as an issue on
+`dkackman/diffusers-workflow` with the `field-report` label, for example:
+
+> File each bug or gap you hit as an issue on dkackman/diffusers-workflow
+> with the label `field-report`.
+
+The label marks a report as coming from real use, not from the automated test
+loop. The agent loop (see [Agent Loop](docs/AGENT_LOOP.md)) picks the report up
+like any other issue when you filed it yourself; a report filed under any
+other GitHub login is parked for the maintainer to review first (relabelled
+`owner:don` + `status:needs-approval`), since the loop must not act
+unattended on third-party text in a public repo, and is only handed to the
+loop, or not, after that review. The label is also what feature planning
+reads as evidence of demand.
+
 The complete tool reference, client configuration for other MCP hosts, and the
 troubleshooting table: [MCP Server](docs/MCP.md). Workspaces in depth:
 [Workspaces](docs/WORKSPACES.md).

{diffusers_workflow-0.4.0b7 → diffusers_workflow-0.5.0/diffusers_workflow.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffusers-workflow
-Version: 0.4.0b7
+Version: 0.5.0
 Summary: Declarative workflow engine and web UI for Hugging Face Diffusers
 Author: dkackman
 License: Apache-2.0
@@ -47,7 +47,10 @@ Requires-Dist: imageio-ffmpeg>=0.6.0
 Requires-Dist: av>=18.1.0
 Requires-Dist: soundfile>=0.14.0
 Requires-Dist: matplotlib>=3.11.1
+Requires-Dist: pyloudnorm>=0.2.0
+Requires-Dist: scipy>=1.18.1
 Requires-Dist: speechbrain>=1.0.3
+Requires-Dist: demucs>=4.1.0
 Requires-Dist: qrcode>=8.2
 Requires-Dist: requests>=2.34.2
 Requires-Dist: aiohttp>=3.14.3
@@ -109,8 +112,8 @@ source ./activate
 .\venv\scripts\activate
 ```
 
-`python -m dw.test` confirms torch and diffusers import and shows which
-accelerator was found.
+`python -m dw.test` runs a small built-in workflow end to end - it downloads
+SD 1.5 (a few GB) and generates one image on whichever accelerator was found.
 
 **2. Start the engine.** Leave it running; everything else talks to it.
 
@@ -160,7 +163,7 @@ once; without it the run fails partway through with a 401/403 from the Hub.
 
 ## Drive it from an agent
 
-Then just ask. The agent has 58 tools covering the whole surface — the
+Then just ask. The agent has 59 tools covering the whole surface — the
 workflow catalog, the real diffusers pipeline signatures, the job queue, the
 gallery, the model cache:
 
@@ -199,6 +202,23 @@ and outputs — so two agents, or an agent and you in the browser, share the GPU
 without saving over each other. An agent calls `use_workspace` once and the
 rest of the session lands there.
 
+**Feedback from a session.** At the end of a working session, ask the agent
+what got in its way: bugs, gaps, misleading skill text, tools it reached for
+and couldn't find. Have it file each one as an issue on
+`dkackman/diffusers-workflow` with the `field-report` label, for example:
+
+> File each bug or gap you hit as an issue on dkackman/diffusers-workflow
+> with the label `field-report`.
+
+The label marks a report as coming from real use, not from the automated test
+loop. The agent loop (see [Agent Loop](docs/AGENT_LOOP.md)) picks the report up
+like any other issue when you filed it yourself; a report filed under any
+other GitHub login is parked for the maintainer to review first (relabelled
+`owner:don` + `status:needs-approval`), since the loop must not act
+unattended on third-party text in a public repo, and is only handed to the
+loop, or not, after that review. The label is also what feature planning
+reads as evidence of demand.
+
 The complete tool reference, client configuration for other MCP hosts, and the
 troubleshooting table: [MCP Server](docs/MCP.md). Workspaces in depth:
 [Workspaces](docs/WORKSPACES.md).

{diffusers_workflow-0.4.0b7 → diffusers_workflow-0.5.0}/diffusers_workflow.egg-info/SOURCES.txt

@@ -11,10 +11,13 @@ diffusers_workflow.egg-info/top_level.txt
 dw/__init__.py
 dw/adapter_compatibility.py
 dw/arguments.py
+dw/assessment_rules.py
 dw/assets.py
 dw/cache_blocks.json
 dw/cache_blocks.py
 dw/content_types.py
+dw/dissolve_frame_errors.py
+dw/download_watch.py
 dw/elision.py
 dw/events.py
 dw/for_each.py
@@ -25,11 +28,14 @@ dw/introspection.py
 dw/kernel_availability.py
 dw/locations.py
 dw/log_setup.py
+dw/loudness.py
 dw/media_audio.py
 dw/media_frames.py
 dw/media_info.py
+dw/null_media.py
 dw/plan.py
 dw/previous_results.py
+dw/probe_paths.py
 dw/prompt_schema.json
 dw/prompt_weighting.py
 dw/prompts.py
@@ -40,6 +46,7 @@ dw/repl.py
 dw/repl_commands.py
 dw/repl_worker.py
 dw/result.py
+dw/result_fps.py
 dw/run.py
 dw/runs.py
 dw/scalar_result_validation.py
@@ -48,6 +55,9 @@ dw/security.py
 dw/select_validation.py
 dw/serve.py
 dw/settings.py
+dw/shot_span_preflight.py
+dw/shots.py
+dw/slice_preflight.py
 dw/step.py
 dw/step_cache.py
 dw/subfolders.py
@@ -59,7 +69,10 @@ dw/type_helpers.py
 dw/validate.py
 dw/variable_constraints.py
 dw/variables.py
+dw/video_extensions.py
+dw/video_size_errors.py
 dw/vram_estimate.py
+dw/vram_inheritance.py
 dw/worker.py
 dw/workflow.py
 dw/workflow_schema.json
@@ -92,6 +105,7 @@ dw/pipeline_processors/pipeline.py
 dw/pipeline_processors/remote.py
 dw/server/__init__.py
 dw/server/app.py
+dw/server/assess.py
 dw/server/catalog_shape.py
 dw/server/enhancers.py
 dw/server/exports.py
@@ -117,45 +131,45 @@ dw/server/ui/assets/csharp-BKxtCVv1.js
 dw/server/ui/assets/csp-bTuwJoIa.js
 dw/server/ui/assets/css-DIMkf-bt.js
 dw/server/ui/assets/css.worker-B3ciXF_0.js
-dw/server/ui/assets/cssMode-xQCqMuag.js
+dw/server/ui/assets/cssMode-DRY1oIAQ.js
 dw/server/ui/assets/cypher-CVaqCwHa.js
 dw/server/ui/assets/dart-onAF5SnQ.js
 dw/server/ui/assets/dockerfile-DZFCIeNp.js
 dw/server/ui/assets/ecl-D05T4iGw.js
 dw/server/ui/assets/editor-jjEx9u7D.css
-dw/server/ui/assets/editor.api-Bfiq9hmW.js
+dw/server/ui/assets/editor.api-C_Q9knYh.js
 dw/server/ui/assets/editor.worker-q-txB4vs.js
 dw/server/ui/assets/elixir-6RTg0lbw.js
 dw/server/ui/assets/flow9-C5_-GSwl.js
-dw/server/ui/assets/freemarker2-BwDnlSuz.js
+dw/server/ui/assets/freemarker2-CoeF3Y-p.js
 dw/server/ui/assets/fsharp-C8Ef5oNN.js
 dw/server/ui/assets/go-C-y9NEjX.js
 dw/server/ui/assets/graphql-fmXr3nnJ.js
-dw/server/ui/assets/handlebars-Beb1we4G.js
+dw/server/ui/assets/handlebars-ClJeebFA.js
 dw/server/ui/assets/hcl-CpzslTdj.js
-dw/server/ui/assets/html-Yw-qdTCH.js
+dw/server/ui/assets/html-Bp7ECOGB.js
 dw/server/ui/assets/html.worker-C93Ht9o9.js
-dw/server/ui/assets/htmlMode-Dbp7eBPO.js
-dw/server/ui/assets/index-CsXMC8a8.css
-dw/server/ui/assets/index-DqbiZNFq.js
+dw/server/ui/assets/htmlMode-DWzbYXCs.js
+dw/server/ui/assets/index-3Vw6WAPW.css
+dw/server/ui/assets/index-Bv4OgazC.js
 dw/server/ui/assets/ini-sBoK_t0W.js
 dw/server/ui/assets/java-BEtHBSE6.js
-dw/server/ui/assets/javascript-BNzP7yQZ.js
+dw/server/ui/assets/javascript-B6Cbtyg8.js
 dw/server/ui/assets/json.worker-B2V3pomh.js
-dw/server/ui/assets/jsonMode-DxIEESAg.js
+dw/server/ui/assets/jsonMode-dg5n9059.js
 dw/server/ui/assets/julia-Bri6UV-V.js
 dw/server/ui/assets/kotlin-BOotOW0E.js
 dw/server/ui/assets/less-B9JPFI3C.js
 dw/server/ui/assets/lexon-CfSJPG6W.js
-dw/server/ui/assets/liquid-CeQTai1N.js
-dw/server/ui/assets/lspLanguageFeatures-DKVB_UCB.js
+dw/server/ui/assets/liquid-CAyPEpMs.js
+dw/server/ui/assets/lspLanguageFeatures-BGj-zQOH.js
 dw/server/ui/assets/lua-CsQS60Ue.js
 dw/server/ui/assets/m3-D-oSqn_W.js
 dw/server/ui/assets/markdown-Cimd5fb3.js
-dw/server/ui/assets/mdx-Bczl_PHr.js
+dw/server/ui/assets/mdx-B5EuIkzP.js
 dw/server/ui/assets/mips-CIPQ_RoX.js
 dw/server/ui/assets/monaco--ixms01u.css
-dw/server/ui/assets/monaco-Bu2OB6wI.js
+dw/server/ui/assets/monaco-ULLHC90E.js
 dw/server/ui/assets/msdax-DauUninz.js
 dw/server/ui/assets/mysql-SOo6toE5.js
 dw/server/ui/assets/objective-c-FvmIjYaQ.js
@@ -170,10 +184,10 @@ dw/server/ui/assets/powerquery-D3hlyOfw.js
 dw/server/ui/assets/powershell-DmHpPYUd.js
 dw/server/ui/assets/protobuf-C531GsRP.js
 dw/server/ui/assets/pug-Z5eAx3Zn.js
-dw/server/ui/assets/python-CeTsj82v.js
+dw/server/ui/assets/python-C5D4i8ni.js
 dw/server/ui/assets/qsharp-DkqhCAOL.js
 dw/server/ui/assets/r-BwWrilGY.js
-dw/server/ui/assets/razor-CN7JaWqc.js
+dw/server/ui/assets/razor-DL9EVLH-.js
 dw/server/ui/assets/redis-ClamHrr6.js
 dw/server/ui/assets/redshift-DT7zqm-g.js
 dw/server/ui/assets/restructuredtext-BYgofb2h.js
@@ -193,15 +207,16 @@ dw/server/ui/assets/swift-Bxkupp3x.js
 dw/server/ui/assets/systemverilog-Bz4Y3fRF.js
 dw/server/ui/assets/tcl-DISqw1ZD.js
 dw/server/ui/assets/ts.worker-D7T1-Ig5.js
-dw/server/ui/assets/tsMode-CoSEMx-L.js
+dw/server/ui/assets/tsMode-BtDRLqfW.js
 dw/server/ui/assets/twig-De2hgUGE.js
-dw/server/ui/assets/typescript-CcsolRTO.js
+dw/server/ui/assets/typescript-DYpkXm7y.js
 dw/server/ui/assets/typespec-B8J7ngcE.js
 dw/server/ui/assets/vb-DV3o63ZY.js
 dw/server/ui/assets/wgsl-DpFanUEy.js
-dw/server/ui/assets/workers-DLaycdy4.js
-dw/server/ui/assets/xml-QEwuyn1w.js
-dw/server/ui/assets/yaml-BB4U5D-R.js
+dw/server/ui/assets/workers-B4R84wCs.js
+dw/server/ui/assets/xml-CQNiPADm.js
+dw/server/ui/assets/yaml-iWBPUkB-.js
+dw/tasks/assess.py
 dw/tasks/audio_transcription.py
 dw/tasks/audio_utils.py
 dw/tasks/background_remover.py
@@ -213,6 +228,7 @@ dw/tasks/diffusion_upscale.py
 dw/tasks/dissolve_videos.py
 dw/tasks/format_messages.py
 dw/tasks/gather.py
+dw/tasks/grade.py
 dw/tasks/image_to_text.py
 dw/tasks/image_utils.py
 dw/tasks/interpolate_frames.py
@@ -232,6 +248,7 @@ dw/tasks/text_generation.py
 dw/tasks/text_sections.py
 dw/tasks/upscale.py
 dw/tasks/video_utils.py
+dw/tasks/voice_attribution.py
 dw/tasks/zoe_depth.py
 dw/workflows/h3_context_ir.json
 dw/workflows/test.json
@@ -251,7 +268,11 @@ dw_mcp/server.py
 dw_mcp/workspaces.py
 tests/test_argument_updates.py
 tests/test_arguments.py
+tests/test_assemble_and_score_target_lufs.py
+tests/test_assess.py
+tests/test_assessment_rules.py
 tests/test_assets.py
+tests/test_attention_slicing.py
 tests/test_audio_utils.py
 tests/test_best_of_n_template.py
 tests/test_borders.py
@@ -260,7 +281,9 @@ tests/test_catalog_shape.py
 tests/test_catalog_structure.py
 tests/test_chain.py
 tests/test_cli.py
+tests/test_component_name_errors.py
 tests/test_component_pruning.py
+tests/test_component_type_errors.py
 tests/test_compose_text.py
 tests/test_concat_videos.py
 tests/test_config_objects.py
@@ -272,16 +295,20 @@ tests/test_device_helpers.py
 tests/test_device_portability.py
 tests/test_dialogue_short_voices.py
 tests/test_diffusion_upscale.py
+tests/test_dissolve_frame_errors.py
 tests/test_dissolve_videos.py
 tests/test_docs_links.py
+tests/test_download_watch.py
 tests/test_elision.py
 tests/test_events.py
 tests/test_examples.py
 tests/test_for_each.py
 tests/test_gather.py
+tests/test_grade.py
 tests/test_h3_adapters.py
 tests/test_h3_context_ir.py
 tests/test_h3_schedule.py
+tests/test_h3_vram_ceiling.py
 tests/test_hf_placement.py
 tests/test_host_memory.py
 tests/test_host_memory_projection.py
@@ -302,7 +329,9 @@ tests/test_library_sources.py
 tests/test_list_images.py
 tests/test_locations.py
 tests/test_lora_alpha.py
+tests/test_lora_disable.py
 tests/test_ltx2_diffusion_decode.py
+tests/test_ltx2_extend_clip.py
 tests/test_ltx2_ic_loras.py
 tests/test_ltx_prompt_library.py
 tests/test_mcp_assets.py
@@ -323,13 +352,18 @@ tests/test_media_frames.py
 tests/test_media_info.py
 tests/test_mix_audio.py
 tests/test_model_cache.py
+tests/test_modular_output_properties.py
 tests/test_modular_pipeline.py
 tests/test_modular_progress.py
+tests/test_mps_fallback_warning.py
+tests/test_mps_rope_precision.py
+tests/test_null_media.py
 tests/test_observed_cost.py
 tests/test_offload_placement.py
 tests/test_output_references.py
 tests/test_pair_audio.py
 tests/test_pair_audio_fit.py
+tests/test_parallel_loading_default.py
 tests/test_phase_events.py
 tests/test_pipeline_caching.py
 tests/test_pipeline_components.py
@@ -347,23 +381,30 @@ tests/test_reference_limits.py
 tests/test_reference_names.py
 tests/test_repl_commands.py
 tests/test_repl_ergonomics.py
-tests/test_repl_hierarchical.py
-tests/test_repl_reorganization.py
 tests/test_repl_worker.py
 tests/test_rerun_new_seed.py
 tests/test_resize_bucket.py
 tests/test_result.py
+tests/test_result_fps.py
 tests/test_result_output_naming.py
 tests/test_runs.py
 tests/test_scalar_result_validation.py
 tests/test_scheduler_config.py
 tests/test_schema.py
 tests/test_security.py
+tests/test_security_auth.py
+tests/test_security_decoder_bombs.py
+tests/test_security_input_caps.py
+tests/test_security_ssrf.py
+tests/test_security_symlinks.py
+tests/test_security_trust_gate.py
+tests/test_security_web_headers.py
 tests/test_segment.py
 tests/test_select.py
 tests/test_select_validation.py
 tests/test_serve_main.py
 tests/test_server.py
+tests/test_server_assess.py
 tests/test_server_downloads.py
 tests/test_server_exports.py
 tests/test_server_guides.py
@@ -371,6 +412,9 @@ tests/test_server_info.py
 tests/test_server_jobs.py
 tests/test_server_mcp.py
 tests/test_server_workspaces.py
+tests/test_shot_span_preflight.py
+tests/test_shots.py
+tests/test_slice_preflight.py
 tests/test_speech_generation.py
 tests/test_stabilize.py
 tests/test_step.py
@@ -382,6 +426,7 @@ tests/test_task_discovery.py
 tests/test_task_domains.py
 tests/test_task_signature_errors.py
 tests/test_teacache.py
+tests/test_template_defaults.py
 tests/test_template_subfolders.py
 tests/test_tensor_image.py
 tests/test_text_generation.py
@@ -392,9 +437,15 @@ tests/test_updater.py
 tests/test_validate_arguments.py
 tests/test_variable_constraints.py
 tests/test_variables.py
+tests/test_video_extensions.py
+tests/test_video_size_errors.py
 tests/test_video_utils.py
+tests/test_voice_attribution.py
+tests/test_vram_estimate.py
+tests/test_vram_inheritance.py
 tests/test_worker.py
 tests/test_worker_execute.py
+tests/test_worker_job_peak_rss.py
 tests/test_workflow.py
 tests/test_workflow_sources.py
 tests/test_workflow_step_cache.py

{diffusers_workflow-0.4.0b7 → diffusers_workflow-0.5.0}/diffusers_workflow.egg-info/requires.txt

@@ -28,7 +28,10 @@ imageio-ffmpeg>=0.6.0
 av>=18.1.0
 soundfile>=0.14.0
 matplotlib>=3.11.1
+pyloudnorm>=0.2.0
+scipy>=1.18.1
 speechbrain>=1.0.3
+demucs>=4.1.0
 qrcode>=8.2
 requests>=2.34.2
 aiohttp>=3.14.3

{diffusers_workflow-0.4.0b7 → diffusers_workflow-0.5.0}/dw/__init__.py

@@ -1,8 +1,11 @@
 from .settings import resolve_path, load_settings
 from .log_setup import setup_logging
+import functools
 import os
 import re
 import logging
+import subprocess
+import sys
 import warnings
 from dotenv import load_dotenv
 
@@ -20,14 +23,35 @@ if "PYTORCH_MPS_HIGH_WATERMARK_RATIO" not in os.environ:
 if "PYTORCH_CUDA_ALLOC_CONF" not in os.environ:
     os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
 
-# Load sharded checkpoints in parallel - a pure cold-start win
+
+def _parallel_loading_default(platform):
+    """Whether diffusers loads a sharded checkpoint on several threads.
+
+    A pure cold-start win on CUDA. On MPS the loader threads each copy their
+    tensors onto the device at once, and that segfaulted LTX-2.5's SDNQ
+    transformer load (exit 139, four threads inside copy_ to mps) where the
+    serial load succeeded in 39s. diffusers reads the variable once at import,
+    before dw can ask torch for a device, so this goes by platform: macOS has
+    no CUDA, and its accelerator is always MPS."""
+    return "false" if platform == "darwin" else "true"
+
+
 if "HF_ENABLE_PARALLEL_LOADING" not in os.environ:
-    os.environ["HF_ENABLE_PARALLEL_LOADING"] = "true"
+    os.environ["HF_ENABLE_PARALLEL_LOADING"] = _parallel_loading_default(sys.platform)
 
 # Suppress all common library warnings before any imports
 warnings.filterwarnings("ignore", category=FutureWarning)
 warnings.filterwarnings("ignore", category=UserWarning)
 warnings.filterwarnings("ignore", category=DeprecationWarning)
+# ...except torch's notice that an op the MPS backend lacks ran on the CPU. With
+# PYTORCH_ENABLE_MPS_FALLBACK set (the fp4-fp8-for-torch-mps autoload sets it)
+# that is otherwise a step that got several times slower with nothing to say why.
+# Later filters take precedence, so this one wins over the blanket ignore above
+warnings.filterwarnings(
+    "default",
+    message=r".*not currently supported on the MPS backend and will fall back",
+    category=UserWarning,
+)
 
 # Eager (but guarded) import of torch. This module's own top-level code needs
 # torch's exceptions/types unconditionally to detect its absence gracefully
@@ -299,10 +323,11 @@ def device_memory_stats():
     """
     Snapshot of allocator memory usage for the device dw is configured for.
 
-    Only CUDA exposes free/total figures (via torch.cuda.mem_get_info); MPS
-    has no equivalent API, so its snapshot reports zeroed, "known" values
-    instead. CPU, or a CUDA/MPS backend that isn't actually available despite
-    being the configured device type, results in an "unavailable" snapshot.
+    CUDA exposes free/total via torch.cuda.mem_get_info. MPS reports
+    allocated (tensors), reserved (everything the Metal driver holds) and
+    total (Metal's recommended working set); free is total minus reserved.
+    CPU, or a CUDA/MPS backend that isn't actually available despite being
+    the configured device type, results in an "unavailable" snapshot.
 
     Returns:
         dict with keys:
@@ -310,10 +335,10 @@ def device_memory_stats():
             device_name (str or None)
             allocated_mb (float)
             reserved_mb (float)
-            free_mb (float or None): None only when CUDA's mem_get_info call
-                itself fails
-            total_mb (float or None): None only when CUDA's mem_get_info call
-                itself fails
+            free_mb (float or None): None only when the backend's capacity
+                probe itself fails
+            total_mb (float or None): None only when the backend's capacity
+                probe itself fails
     """
     stats = {
         "available": False,
@@ -344,15 +369,59 @@ def device_memory_stats():
         and hasattr(torch.backends, "mps")
         and torch.backends.mps.is_available()
     ):
-        # MPS doesn't provide detailed memory stats like CUDA
+        # Unified memory: 'reserved' is everything the Metal driver holds for
+        # this process, and the ceiling is Metal's recommended working set -
+        # the figure past which an allocation starts to page
         stats["available"] = True
-        stats["device_name"] = "Apple Silicon (MPS)"
-        stats["free_mb"] = 0.0
-        stats["total_mb"] = 0.0
+        stats["device_name"] = f"{_apple_chip_name()} (MPS)"
+        stats["allocated_mb"] = torch.mps.current_allocated_memory() / 1024 / 1024
+        stats["reserved_mb"] = torch.mps.driver_allocated_memory() / 1024 / 1024
+        try:
+            total = torch.mps.recommended_max_memory() / 1024 / 1024
+            stats["total_mb"] = total
+            stats["free_mb"] = max(total - stats["reserved_mb"], 0.0)
+        except (RuntimeError, AttributeError):
+            pass
 
     return stats
 
 
+@functools.lru_cache(maxsize=1)
+def _apple_chip_name():
+    """The chip's marketing name ('Apple M5 Pro'), so a reading from an 8 GB M1
+    and one from a 128 GB M4 Max are not reported as the same device."""
+    try:
+        result = subprocess.run(
+            ["sysctl", "-n", "machdep.cpu.brand_string"],
+            capture_output=True,
+            text=True,
+            timeout=2,
+            check=True,
+        )
+        return result.stdout.strip() or "Apple Silicon"
+    except (OSError, subprocess.SubprocessError):
+        return "Apple Silicon"
+
+
+def device_capacity_gb(device=None):
+    """How much memory the accelerator can hold, in GiB - Metal's recommended
+    working set on MPS, the card's total on CUDA - or None where there is no
+    accelerator or the probe fails. Used where a check needs a ceiling for a
+    device no curated 'cost' entry describes."""
+    if not _TORCH_AVAILABLE:
+        return None
+    try:
+        device_type = get_device_type(device)
+        if device_type == "mps":
+            return torch.mps.recommended_max_memory() / 1024**3
+        if device_type == "cuda":
+            index = torch.device(device or get_device()).index or 0
+            return torch.cuda.get_device_properties(index).total_memory / 1024**3
+    except (RuntimeError, AttributeError, AssertionError):
+        return None
+    return None
+
+
 def startup(log_level=None):
     if not _TORCH_AVAILABLE:
         raise ImportError(